The Abstraction and Reasoning Corpus (ARC) is a dataset that measures general fluid intelligence in AI systems. It consists of tasks where the AI must infer a pattern from a few examples and apply it to new situations.
Each task contains:
This page showcases different transduction models attempting to solve the ARC validation set. For each task, models generate multiple candidate solutions, which are ranked based on various strategies including test-time fine-tuning and reranking approaches.
The visualization allows you to:
For implementation details about the models and evaluation process, visit: github.com/xu3kev/BARC/blob/master/seeds/common.py